betaFeedback

WorksComputer Science

S-R3: Secure and Robust RIS-Assisted RSMA under Finite-Resolution Phase Shifts

Abstract

This paper investigates a reconfigurable intelligent surface (RIS)-assisted rate-splitting multiple access (RSMA) downlink with imperfect eavesdropper channel state information and finite-resolution RIS phases. The common stream carries public information, while user-specific private streams carry confidential information. For a conservative security assessment, each eavesdropper is allowed to remove the public stream and is given the other private streams as side information. We formulate a robust design of t he transmit precoders, isotropicartificial noise, common-rate allocations, and RIS phases that balances confidential and public throughput under rate and power constraints. Bounded errors in the direct and cascaded eavesdropper channels yield an effective-channel uncertainty bound valid across RIS phase updates. To address the resulting nonconvex problem, we construct conservative leakage bounds and convex inner transmit updates, initialize through a feasibility procedure, and search finite artificial-noise and RIS-phase candidates with explicit feasibility checks. For a fixed artificialnoise power and RIS configuration, the feasible inner iterates have a nondecreasing objective that converges in value, and their successive changes vanish. Accepted outer updates remain feasible and are finite in number. Our code is available at https://github.com/ABM-ADAM/RIS RSMA-secure.

Introduction

Reconfigurable intelligent surfaces (RISs) shape reflected propagation paths in a multiuser downlink, affecting the channels of both legitimate users and potential eavesdroppers. Secure RIS-assisted transmission therefore calls for jointly designing the base-station precoders, artificial noise (AN), and RIS phases [1], [2]. Practical RIS elements provide a finite set of phase states, whereas continuous-phase control is an idealization [3]. Furthermore, the eavesdropper channels may be imperfectly known at the transmitter [1]. A security formulation should account for both finite phase resolution and eavesdropper-channel uncertainty.

Rate-splitting multiple access (RSMA) splits user messages into common and private parts: the common stream is decoded by all served users before each user decodes its own private stream. This structure can also accommodate multicast and unicast traffic in the same transmission [4]. We assign different security requirements to the two streams: the common stream carries public message parts, whereas each private stream carries the confidential part of one user’s message. Thus, the public common rate contributes to delivered throughput but never to confidential throughput. Unlike RIS–RSMA formulations that allocate a confidential common rate [5], our security model allows an eavesdropper to know and cancel the public common stream. The RIS, transmit precoders, and AN are consequently designed to protect each confidential private stream under that assumption.

This paper studies that mixed-traffic downlink with bounded eavesdropper-channel errors and finite-resolution RIS phases. The objective balances a weighted sum of guaranteed confidential private rates against delivered public common rate, subject to explicit minimum-rate and transmit-power constraints. Because the resulting problem couples robust leakage bounds, multiuser decoding, AN, and discrete phase choices, both the security meaning of a rate and the scope of an algorithmic guarantee require careful specification.

RIS-assisted secure transmission has been investigated in several aerial and multiuser settings. Joint transmit beamforming, RIS phases, and unmanned aerial vehicle (UAV) trajectory are optimized in [6], while the impact of UAV jitter and imperfect hardware is addressed in [7]. More recently, hybrid RISs and multihop propagation have been studied for secure indoor–outdoor aerial millimeter-wave links [8]; secure content fetching in the presence of aerial and ground eavesdroppers is considered in [9]. These settings establish the relevance of propagation control and uncertain operating conditions, although their airborne geometry and traffic objectives differ from the fixed, narrowband downlink considered here.

Secure RSMA has also been studied in UAV networks through deep unfolding and reinforcement learning [10], quality-of-experience-aware resource allocation [11], and quantum deep reinforcement learning [12]. For RIS-assisted multiuser secrecy, the precoding study in [13] develops an alternating-optimization solution and an unfolded network for its original model. The closer RIS–RSMA work of Wang et al. [5] optimizes a max–min secrecy objective with transceiver hardware impairments and an allocation of confidential common rate. In contrast, the present common stream is public and may be known to the eavesdroppers. Its allocation cannot be inserted into a secrecy sum, and a comparison with [5] requires adapting the formulations to the same traffic and channel assumptions.

Learning-based methods address related but distinct RIS control problems: hierarchical reinforcement learning for terahertz massive MIMO [14], multimodal reinforcement learning for an interference-limited RIS–UAV network [15], and a discrete policy-gradient method for statistical quality of service in RIS-aided holographic MIMO [16]. A quantum-enhanced multi-RIS approach has also been studied for millimeter-wave Internet-of-Things networks using nonorthogonal multiple access [17]. These works illustrate different channel, traffic, or control models; none of their reported performance numbers is used as a surrogate for the robust public-common/private-confidential problem posed here.

Motivations and Contributions

The above literature motivates a formulation in which the security claim matches the information actually required to be secret. Counting a public common allocation as secrecy, or relying on other users’ private streams to remain undecodable at an eavesdropper, would overstate the confidentiality provided by this model. A second issue is that RIS phase updates change the effective eavesdropper channel, so uncertainty defined only around one fixed RIS configuration need not remain valid after an update. Finally, finite phase resolution and coupled rate constraints call for an implementable procedure with guarantees stated for the problem it actually solves. The contributions are as follows:

  • A traffic-consistent robust model: We formulate a downlink with a public common stream, confidential private streams, isotropic AN, and a finite-resolution passive RIS. Eavesdroppers are allowed to cancel the public stream and, conservatively, are given the other private streams as side information. We bound direct-link and cascaded-channel estimation errors and derive an enclosing effective-channel uncertainty ball valid for every admissible RIS configuration. The public and confidential-rate constraints and their objective weights remain separate.

  • A verifiable joint design procedure: We upper-bound worst-case private-stream leakage and construct convex inner transmit updates for fixed AN power and RIS phases using rate minorization, leakage majorization, and a proximal term. A Phase I procedure searches for a feasible starting point; finite AN-power candidates and quantized RIS coordinate changes are then evaluated with direct checks of the physical constraints before acceptance. The method returns a feasible conservative design whenever its initialization and subsequent numerical solves succeed.

  • Explicit guarantees and evaluation protocol: We provide pseudocode and arithmetic-complexity scaling for the finite-budget procedure. For a fixed AN-power/RIS pair and a feasible initial point, the inner objective is monotone and converges in value, with successive steps vanishing under the stated proximal update. Accepted outer moves remain feasible and improve the objective by a prescribed positive threshold, so their number is finite. The numerical methodology separates confidential and public rates and calls for an adapted RIS–RSMA @bibliography

literature baseline evaluated on paired channel draws. Section II specifies the signal model and robust problem. Section III develops the solution and its guarantees, and Section IV describes the simulation methodology and results.

System Model and Problem Formulation

Network and Channel Model

We consider the downlink of a reconfigurable intelligent surface (RIS)-assisted rate-splitting multiple access (RSMA) network, as depicted in Figure 1. A base station (BS) equipped with NN antennas serves KK single-antenna legitimate users in the presence of EE noncolluding, single-antenna eavesdroppers. The RIS has MM passive reflecting elements. The sets of legitimate users and eavesdroppers are denoted by K={1,…,K}\mathcal{K}=\{1,\ldots,K\} and E={1,…,E}\mathcal{E}=\{1,\ldots,E\}, respectively.

Figure 1. RIS-assisted RSMA downlink with noncolluding eavesdroppers. ChatGPT-6 Astra is used to plot the figure.

The BS-to-RIS channel is G∈CM×N\mathbf{G}\in\mathbb{C}^{M\times N}. For legitimate user kk, let hd,k∈CN\mathbf{h}_{d,k}\in\mathbb{C}^{N} and hr,k∈CM\mathbf{h}_{r,k}\in\mathbb{C}^{M} denote the direct BS-to-user and RIS-to-user channels, respectively. Similarly, hd,e∈CN\mathbf{h}_{d,e}\in\mathbb{C}^{N} and hr,e∈CM\mathbf{h}_{r,e}\in\mathbb{C}^{M} denote the channels associated with eavesdropper ee. A blocked direct link is represented by setting its corresponding direct-channel vector to zero; it is not assigned a nonzero line-of-sight component.

The RIS employs BB-bit phase control with alphabet

QB={ej2πℓ/2B:ℓ=0,…,2B−1}.(1)\mathcal{Q}_{B}=\left\{e^{\mathrm{j}2\pi\ell/2^{B}}:\ell=0,\ldots,2^{B}-1\right\}. \tag*{(1)}

The RIS reflection coefficients are collected in

θ=[ejϕ1,…,ejϕM]T,\boldsymbol{\theta}=\left[e^{\mathrm{j}\phi_{1}},\ldots,e^{\mathrm{j}\phi_{M}}\right]^{T},
Θ=diag⁡(θ),θm∈QB.(2)\boldsymbol{\Theta}=\operatorname{diag}(\boldsymbol{\theta}),\quad\theta_{m}\in\mathcal{Q}_{B}. \tag*{(2)}

For a fixed RIS configuration, the effective BS-to-receiver channels are

hkH(θ)=hd,kH+hr,kHΘG,k∈K,(3)\mathbf{h}_{k}^{H}(\boldsymbol{\theta})=\mathbf{h}_{d,k}^{H}+\mathbf{h}_{r,k}^{H}\boldsymbol{\Theta}\mathbf{G},\qquad k\in\mathcal{K}, \tag*{(3)}
heH(θ)=hd,eH+hr,eHΘG,e∈E.(4)\mathbf{h}_{e}^{H}(\boldsymbol{\theta})=\mathbf{h}_{d,e}^{H}+\mathbf{h}_{r,e}^{H}\boldsymbol{\Theta}\mathbf{G},\qquad e\in\mathcal{E}. \tag*{(4)}

Both effective channels in (3)–(4) have dimension 1×N1\times N. The channel generation model, including path loss, fading distributions, spatial correlation, and any line-of-sight components, will be specified separately in the simulation methodology. These statistical assumptions are not needed to define the instantaneous signal and optimization models below.

We assume that the BS and RIS controller know the legitimate-user channels during the considered transmission interval. For eavesdropper ee, write its effective-channel row vector as

heH(θ)=hd,eH+θTBe,Be=diag⁡(hr,e∗)G∈CM×N.(5)\mathbf{h}_{e}^{H}(\boldsymbol{\theta})=\mathbf{h}_{d,e}^{H}+\boldsymbol{\theta}^{T}\mathbf{B}_{e},\qquad\mathbf{B}_{e}=\operatorname{diag}(\mathbf{h}_{r,e}^{*})\mathbf{G}\in\mathbb{C}^{M\times N}. \tag*{(5)}

The controller has estimates h^d,e\widehat{\mathbf{h}}_{d,e} and B^e\widehat{\mathbf{B}}_{e}, while the underlying channel errors obey

∥hd,e−h^d,e∥2≤ϵd,e,∥Be−B^e∥F≤ϵB,e.(6)\left\|\mathbf{h}_{d,e}-\widehat{\mathbf{h}}_{d,e}\right\|_{2}\leq\epsilon_{d,e},\qquad\left\|\mathbf{B}_{e}-\widehat{\mathbf{B}}_{e}\right\|_{F}\leq\epsilon_{B,e}. \tag*{(6)}

Define the estimated effective channel by h^eH(θ)=h^d,eH+θTB^e\widehat{\mathbf{h}}_{e}^{H}(\boldsymbol{\theta})=\widehat{\mathbf{h}}_{d,e}^{H}+\boldsymbol{\theta}^{T}\widehat{\mathbf{B}}_{e}. For every admissible θ\boldsymbol{\theta}, the effective-channel error satisfies

∥he(θ)−h^e(θ)∥2≤ϵe≜ϵd,e+MϵB,e.(7)\left\|\mathbf{h}_{e}(\boldsymbol{\theta})-\widehat{\mathbf{h}}_{e}(\boldsymbol{\theta})\right\|_{2}\leq\epsilon_{e}\triangleq\epsilon_{d,e}+\sqrt{M}\epsilon_{B,e}. \tag*{(7)}

Indeed, ∥θ∥2=M\|\boldsymbol{\theta}\|_{2}=\sqrt{M} and ∥θT(Be−B^e)∥2≤MϵB,e\|\boldsymbol{\theta}^{T}(\mathbf{B}_{e}-\widehat{\mathbf{B}}_{e})\|_{2}\leq\sqrt{M}\epsilon_{B,e}. This bound holds uniformly when the RIS phases change. For compactness, define the enclosing uncertainty set

Ue≜{Δ∈CN:∥Δ∥2≤ϵe},he(θ)=h^e(θ)+Δe.(8)\mathcal{U}_{e}\triangleq\left\{\boldsymbol{\Delta}\in\mathbb{C}^{N}:\|\boldsymbol{\Delta}\|_{2}\leq\epsilon_{e}\right\},\qquad\mathbf{h}_{e}(\boldsymbol{\theta})=\widehat{\mathbf{h}}_{e}(\boldsymbol{\theta})+\boldsymbol{\Delta}_{e}. \tag*{(8)}

The ball Ue\mathcal{U}_{e} can contain effective-channel errors that cannot arise from the primitive errors in (6); protecting against the entire ball is conservative. These radii are deterministic bounds, not Gaussian standard deviations. Setting both radii to zero gives perfect estimated eavesdropper CSI.

RSMA Transmission and Legitimate-User Rates

Each user message is split into a public part and a confidential part. The public parts are jointly encoded into one common stream sc\mathbf{s}_{c}, which every legitimate user decodes. The confidential part of user kk is encoded into its private stream sks_{k}. The streams are independent and satisfy E[∣sc∣2]=E[∣sk∣2]=1\mathbb{E}[|s_{c}|^{2}]=\mathbb{E}[|s_{k}|^{2}]=1. The BS transmits

x=wcsc+∑k∈Kwksk+z,(9)\mathbf{x}=\mathbf{w}_{c}s_{c}+\sum_{k\in\mathcal{K}}\mathbf{w}_{k}s_{k}+\mathbf{z}, \tag*{(9)}

where wc,wk∈CN\mathbf{w}_{c},\mathbf{w}_{k}\in\mathbb{C}^{N} are the common and private precoders. The independently generated artificial noise (AN) obeys z∼CN(0,qIN)\mathbf{z}\sim\mathcal{CN}(\mathbf{0},q\mathbf{I}_{N}), where q≥0q\geq0 is an optimized scalar. Consequently, the total average transmit-power constraint is

∥wc∥22+∑k∈K∥wk∥22+Nq≤PB.(10)\|\mathbf{w}_{c}\|_{2}^{2}+\sum_{k\in\mathcal{K}}\|\mathbf{w}_{k}\|_{2}^{2}+Nq\leq P_{B}. \tag*{(10)}

The AN in (9) is a controllable transmission variable; it does not represent transmitter hardware distortion.

The signal received by legitimate user kk is

yk=hkH(θ)wcsc+∑j∈KhkH(θ)wjsj+hkH(θ)z+nk,(11)y_{k}=\mathbf{h}_{k}^{H}(\boldsymbol{\theta})\mathbf{w}_{c}s_{c}+\sum_{j\in\mathcal{K}}\mathbf{h}_{k}^{H}(\boldsymbol{\theta})\mathbf{w}_{j}s_{j}+\mathbf{h}_{k}^{H}(\boldsymbol{\theta})\mathbf{z}+n_{k}, \tag*{(11)}

where nk∼CN(0,σk2)n_{k}\sim\mathcal{CN}(0,\sigma_{k}^{2}) and σk2>0\sigma_{k}^{2}>0. For compactness, define

Dk(θ,q)=q∥hk(θ)∥22+σk2.(12)D_{k}(\boldsymbol{\theta},q)=q\left\|\mathbf{h}_{k}(\boldsymbol{\theta})\right\|_{2}^{2}+\sigma_{k}^{2}. \tag*{(12)}

Each legitimate user first decodes scs_{c} while treating all private streams and AN as noise. Its achievable common-stream rate is

Rc,k=log⁡2(1+∣hkH(θ)wc∣2∑j∈K∣hkH(θ)wj∣2+Dk(θ,q)).(13)R_{c,k}=\log_{2}\left(1+\frac{\left|\mathbf{h}_{k}^{H}(\boldsymbol{\theta})\mathbf{w}_{c}\right|^{2}}{\sum_{j\in\mathcal{K}}\left|\mathbf{h}_{k}^{H}(\boldsymbol{\theta})\mathbf{w}_{j}\right|^{2}+D_{k}(\boldsymbol{\theta},q)}\right). \tag*{(13)}

After decoding and removing scs_{c}, user kk decodes sks_{k}, treating the other private streams and AN as noise. Its achievable private-stream rate is

Rp,k=log⁡2(1+γp,k),(14)R_{p,k}=\log_{2}(1+\gamma_{p,k}), \tag*{(14)}
γp,k=∣hkH(θ)wk∣2∑j≠k∣hkH(θ)wj∣2+Dk(θ,q).(15)\gamma_{p,k}=\frac{\left|\mathbf{h}_{k}^{H}(\boldsymbol{\theta})\mathbf{w}_{k}\right|^{2}}{\sum_{j\neq k}\left|\mathbf{h}_{k}^{H}(\boldsymbol{\theta})\mathbf{w}_{j}\right|^{2}+D_{k}(\boldsymbol{\theta},q)}. \tag*{(15)}

The rates in (13)–(14) assume successful common-stream cancellation. The effect of imperfect successive interference cancellation is outside the present signal model.

Let Ck≥0C_{k}\geq0 be the portion of the common-stream rate assigned to user kk. Because all legitimate users must decode the complete common stream, the common-rate allocation must satisfy

∑k∈KCk≤Rc,i,i∈K.(16)\sum_{k\in\mathcal{K}} C_k \le R_{c,i}, \qquad i\in\mathcal{K}. \tag*{(16)}

Writing the constraint for every ii avoids the nonsmooth min⁡iRc,i\min_i R_{c,i} notation and is equivalent to requiring that the aggregate common rate be decodable by the weakest legitimate user.

Eavesdropper Model and Confidential Rates

The common stream carries public information. We therefore allow an eavesdropper to know and remove scs_c when attempting to decode a confidential private stream. We also give it the other users’ private streams as side information. This latter assumption gives each eavesdropper a stronger decoding capability than simply treating the other private streams as noise; it avoids relying on multiuser interference as the sole source of confidentiality.

For an effective eavesdropper channel he(θ)\mathbf{h}_e(\boldsymbol{\theta}), the resulting information-leakage rate associated with the private stream of user kk is bounded by

Re,k(he(θ))=log⁡2(1+∣heH(θ)wk∣2q∥he(θ)∥2+σe2).(17)R_{e,k}\bigl(\mathbf{h}_e(\boldsymbol{\theta})\bigr) = \log_2\left(1+ \frac{\left|\mathbf{h}_e^{H}(\boldsymbol{\theta})\mathbf{w}_k\right|^2} {q\left\|\mathbf{h}_e(\boldsymbol{\theta})\right\|^2+\sigma_e^2} \right). \tag*{(17)}

where σe2>0\sigma_e^2>0 is the receiver-noise variance of eavesdropper ee. In (17), the other information streams do not appear in the denominator because they are assumed known to the eavesdropper for this conservative leakage calculation. The eavesdroppers are noncolluding: each acts on its own observation.

With independent Gaussian signaling and stochastic wiretap encoding of each confidential private message, the following private-stream rate difference is an achievable confidential-rate bound under the stated noncolluding, genie-aided eavesdropper model:

Rs,k=[Rp,k−max⁡e∈ERe,k(he(θ))]+.(18)R_{s,k} = \left[ R_{p,k}-\max_{e\in\mathcal{E}}R_{e,k}\bigl(\mathbf{h}_e(\boldsymbol{\theta})\bigr) \right]^+. \tag*{(18)}

where [a]+≜max⁡{a,0}[a]^+\triangleq\max\{a,0\}. The public allocation CkC_k is not included in (18). Thus, CkC_k may contribute to a user’s overall delivered rate, but it must not be reported as confidential throughput.

Under the bounded uncertainty model (8), define the worst-case leakage rate

R‾e,k=max⁡e∈E sup⁡Δe∈UeRe,k(h^e(θ)+Δe).(19)\overline{R}_{e,k} = \max_{e\in\mathcal{E}}\ \sup_{\boldsymbol{\Delta}_e\in\mathcal{U}_e} R_{e,k}\bigl(\widehat{\mathbf{h}}_e(\boldsymbol{\theta})+\boldsymbol{\Delta}_e\bigr). \tag*{(19)}

The corresponding robust confidential rate is

Rs,k=[Rp,k−R‾e,k]+.(20)R_{s,k} = \left[R_{p,k}-\overline{R}_{e,k}\right]^+. \tag*{(20)}

The supremum in (19) is taken over a bounded channel-error set. It is not a chance constraint, and no secrecy-outage probability is claimed.

Robust Joint Precoding and RIS Design

Let αk>0\alpha_k>0 denote the confidential-rate weight of user kk, and let λ≥0\lambda\ge0 denote the public-throughput weight. We introduce tk≥0t_k\ge0 as its designed confidential rate. To ensure that the common stream serves an actual traffic requirement, let rkpub≥0r_k^{\mathrm{pub}}\ge0 denote user kk’s minimum public rate. An optional minimum confidential-rate requirement is denoted by rksec≥0r_k^{\mathrm{sec}}\ge0. These are distinct requirements: public bits cannot satisfy a confidential-rate requirement.

The joint public-and-confidential throughput design is parameterized by λ\lambda. We maximize a scalarized objective, while always reporting the public and confidential components separately. For λ=0\lambda=0 the formulation reduces to the confidential-throughput objective; for λ>0\lambda>0 the common stream can carry more than its minimum public traffic requirement. A sweep over λ\lambda traces the attained public–confidential trade-off, subject to the limits of the nonconvex solver. Let V≜(wc,{wk},q,θ,{Ck,tk})\mathcal{V}\triangleq(\mathbf{w}_c,\{\mathbf{w}_k\},q,\boldsymbol{\theta},\{C_k,t_k\}). The design is

max⁡V∑k∈Kαktk+λ∑k∈KCk(21)\begin{aligned} \max_{\mathcal{V}}\quad& \sum_{k\in\mathcal{K}}\alpha_k t_k+\lambda\sum_{k\in\mathcal{K}}C_k \tag*{(21)} \end{aligned}

subject to

tk≤Rp,k−Re,k(h^e(θ)+Δe),∀k∈K, e∈E, Δe∈Ue,(22)t_k\le R_{p,k}-R_{e,k}\bigl(\widehat{\mathbf{h}}_e(\boldsymbol{\theta})+\boldsymbol{\Delta}_e\bigr), \qquad \forall k\in\mathcal{K},\ e\in\mathcal{E},\ \boldsymbol{\Delta}_e\in\mathcal{U}_e, \tag*{(22)}
∑k∈KCk≤Rc,i,∀i∈K,(23)\sum_{k\in\mathcal{K}}C_k\le R_{c,i}, \qquad \forall i\in\mathcal{K}, \tag*{(23)}
Ck≥rkpub,tk≥rksec,∀k∈K,(24)C_k\ge r_k^{\mathrm{pub}},\qquad t_k\ge r_k^{\mathrm{sec}}, \qquad\forall k\in\mathcal{K}, \tag*{(24)}
∥wc∥22+∑k∈K∥wk∥22+Nq≤PB,(25)\left\|\mathbf{w}_c\right\|_2^2+\sum_{k\in\mathcal{K}}\left\|\mathbf{w}_k\right\|_2^2+Nq\le P_B, \tag*{(25)}
q≥0,θm∈QB,m=1,…,M.(26)q\ge0,\qquad\theta_m\in\mathcal{Q}_B,\qquad m=1,\ldots,M. \tag*{(26)}

Constraint (22) limits tkt_k by the legitimate private rate minus the leakage rate for every eavesdropper and every channel in its uncertainty set. Consequently, any feasible tkt_k is no greater than Rs,kR_{s,k} in (20). The term λ∑kCk\lambda\sum_k C_k rewards only public throughput; it is never interpreted as a secrecy rate. Constraints (23) and (24) ensure common-stream decodability and distinguish public from confidential traffic. The remaining constraints impose the BS power budget, nonnegative isotropic AN power, and finite-resolution passive RIS operation.

Problem (20) is nonconvex. In particular, the legitimate and eavesdropper rates are coupled through the precoders, AN covariance, and RIS phases; the semi-infinite secrecy constraints span the channel-uncertainty sets; and the discrete RIS phase decisions are combinatorial. A solution method must therefore specify how it handles each of these features. Neither alternating optimization nor semidefinite relaxation alone establishes global optimality. The next section develops a tractable iterative method and states the guarantees that its individual updates support.

Conservative Robust Transmit and RIS Design

A Uniform Upper Bound on Leakage

For a chosen public-throughput weight λ≥0\lambda\ge0, we seek a feasible, possibly suboptimal solution of (20). For each RIS configuration, define

ae(θ)=[∥h^e(θ)∥2−ϵe]+2,[x]+≜max⁡{x,0}.(27)a_e(\theta)=\left[\lVert\hat{\mathbf{h}}_e(\theta)\rVert_2-\epsilon_e\right]_+^2,\qquad[x]_+\triangleq\max\{x,0\}. \tag*{(27)}

The uniform error bound (7) implies

∣heH(θ)wk∣≤∣h^eH(θ)wk∣+ϵe∥wk∥2,(28)\left|\mathbf{h}_e^H(\theta)\mathbf{w}_k\right|\leq\left|\hat{\mathbf{h}}_e^H(\theta)\mathbf{w}_k\right|+\epsilon_e\lVert\mathbf{w}_k\rVert_2, \tag*{(28)}
∥he(θ)∥22≥ae(θ).(29)\lVert\mathbf{h}_e(\theta)\rVert_2^2\geq a_e(\theta). \tag*{(29)}

Since q≥0q\geq0 and σe2>0\sigma_e^2>0, a uniform upper bound on the leakage rate is Define be,k(θ,wk)≜∣h^eH(θ)wk∣+ϵe∥wk∥2b_{e,k}(\theta,\mathbf{w}_k)\triangleq\left|\hat{\mathbf{h}}_e^H(\theta)\mathbf{w}_k\right|+\epsilon_e\lVert\mathbf{w}_k\rVert_2. Then

R‾e,kub=log⁡2(1+be,k2(θ,wk)σe2+qae(θ)).(30)\overline{R}_{e,k}^{\mathrm{ub}}=\log_2\left(1+\frac{b_{e,k}^2(\theta,\mathbf{w}_k)}{\sigma_e^2+qa_e(\theta)}\right). \tag*{(30)}

For every admissible primitive-channel error, Re,k(he(θ))≤R‾e,kubR_{e,k}(\mathbf{h}_e(\theta))\leq\overline{R}_{e,k}^{\mathrm{ub}}. The numerator and denominator bounds need not be attained by the same channel error. Therefore,

tk+R‾e,kub≤Rp,k,k∈K, e∈E,(31)t_k+\overline{R}_{e,k}^{\mathrm{ub}}\leq R_{p,k},\qquad k\in\mathcal{K},\ e\in\mathcal{E}, \tag*{(31)}

are sufficient, and generally conservative, constraints for (22). In what follows, the conservative physical problem means (20) with its semi-infinite secrecy constraints replaced by (31).

Convex Transmit Update at Fixed AN Power and RIS Phases

Fix q=q0≥0q=q_0\geq0 and θ=θ0\boldsymbol{\theta}=\boldsymbol{\theta}_0. Write hk=hk(θ0)\mathbf{h}_k=\mathbf{h}_k(\boldsymbol{\theta}_0), h^e=h^e(θ0)\hat{\mathbf{h}}_e=\hat{\mathbf{h}}_e(\boldsymbol{\theta}_0), and ae=ae(θ0)a_e=a_e(\boldsymbol{\theta}_0). The precoder-power budget is P0=PB−Nq0P_0=P_B-Nq_0; we exclude pairs with P0<0P_0<0. For given precoders, define

Tc,k=∣hkHwc∣2+∑j∈K∣hkHwj∣2+q0∥hk∥22+σk2,(32)T_{c,k}=\left|\mathbf{h}_k^H\mathbf{w}_c\right|^2+\sum_{j\in\mathcal{K}}\left|\mathbf{h}_k^H\mathbf{w}_j\right|^2+q_0\lVert\mathbf{h}_k\rVert_2^2+\sigma_k^2, \tag*{(32)}
Tp,k=∑j∈K∣hkHwj∣2+q0∥hk∥22+σk2.(33)T_{p,k}=\sum_{j\in\mathcal{K}}\left|\mathbf{h}_k^H\mathbf{w}_j\right|^2+q_0\lVert\mathbf{h}_k\rVert_2^2+\sigma_k^2. \tag*{(33)}

For scalar complex equalizers gc,kg_{c,k} and gp,kg_{p,k}, the common and private mean-square errors (MSEs) are

ec,k=∣gc,k∣2Tc,k−2Re⁡{gc,k∗hkHwc}+1,(34)e_{c,k}=\left|g_{c,k}\right|^2T_{c,k}-2\operatorname{Re}\{g_{c,k}^*\mathbf{h}_k^H\mathbf{w}_c\}+1, \tag*{(34)}
ep,k=∣gp,k∣2Tp,k−2Re⁡{gp,k∗hkHwk}+1.(35)e_{p,k}=\left|g_{p,k}\right|^2T_{p,k}-2\operatorname{Re}\{g_{p,k}^*\mathbf{h}_k^H\mathbf{w}_k\}+1. \tag*{(35)}

For auxiliary MSE weights uc,k,up,k>0u_{c,k},u_{p,k}>0, set

R~c,k=ln⁡uc,k−uc,kec,k+1ln⁡2,(36)\widetilde{R}_{c,k}=\frac{\ln u_{c,k}-u_{c,k}e_{c,k}+1}{\ln2}, \tag*{(36)}
R~p,k=ln⁡up,k−up,kep,k+1ln⁡2.(37)\widetilde{R}_{p,k}=\frac{\ln u_{p,k}-u_{p,k}e_{p,k}+1}{\ln2}. \tag*{(37)}

For arbitrary auxiliaries, R~c,k≤Rc,k\widetilde{R}_{c,k}\leq R_{c,k} and R~p,k≤Rp,k\widetilde{R}_{p,k}\leq R_{p,k}. Both are value tight at the current precoders under the MMSE updates

gc,k=hkHwcTc,k,uc,k=1ec,k,(38)g_{c,k}=\frac{\mathbf{h}_k^H\mathbf{w}_c}{T_{c,k}},\qquad u_{c,k}=\frac{1}{e_{c,k}}, \tag*{(38)}
gp,k=hkHwkTp,k,up,k=1ep,k,(39)g_{p,k}=\frac{\mathbf{h}_k^H\mathbf{w}_k}{T_{p,k}},\qquad u_{p,k}=\frac{1}{e_{p,k}}, \tag*{(39)}

where the MSEs on the right use the corresponding MMSE equalizers. Holding these auxiliaries fixed makes R~c,k\widetilde{R}_{c,k} and R~p,k\widetilde{R}_{p,k} concave quadratic functions of the precoders.

Introduce real ve,k,βe,k≥0v_{e,k},\beta_{e,k}\geq0 and impose

ve,k≥∣h^eHwk∣+ϵe∥wk∥2,(40)v_{e,k}\geq\left|\hat{\mathbf{h}}_e^H\mathbf{w}_k\right|+\epsilon_e\lVert\mathbf{w}_k\rVert_2, \tag*{(40)}
ve,k2≤deβe,k,de≜σe2+q0ae>0.(41)v_{e,k}^2\leq d_e\beta_{e,k},\qquad d_e\triangleq\sigma_e^2+q_0a_e>0. \tag*{(41)}

These convex constraints imply R‾e,kub≤log⁡2(1+βe,k)\overline{R}_{e,k}^{\mathrm{ub}}\leq\log_2(1+\beta_{e,k}). The second inequality has a rotated second-order-cone (SOC) representation because ded_e is constant. If qq were also variable here, that convexity argument would not apply.

At fixed-pair iteration nn, set ve,k(n)v_{e,k}^{(n)} and βe,k(n)\beta_{e,k}^{(n)} to their smallest feasible values at the current precoders. Concavity of log⁡2(1+β)\log_2(1+\beta) gives the affine global upper bound

log⁡2(1+β)≤Le,k(n)(β),\log_2(1+\beta)\leq\mathcal{L}_{e,k}^{(n)}(\beta),
Le,k(n)(β)≜log⁡2(1+βe,k(n))+β−βe,k(n)(1+βe,k(n))ln⁡2.(42)\mathcal{L}_{e,k}^{(n)}(\beta)\triangleq\log_2(1+\beta_{e,k}^{(n)})+\frac{\beta-\beta_{e,k}^{(n)}}{(1+\beta_{e,k}^{(n)})\ln2}. \tag*{(42)}

It is value tight at βe,k(n)\beta_{e,k}^{(n)}. Let X\mathbf{X} collect the real and imaginary parts of all precoders divided by a fixed positive amplitude scale and the real variables {Ck,tk}\{C_k,t_k\} divided by a fixed positive rate scale. This invertible normalization makes the proximal norm dimensionless. Define

F(X)=∑kαktk+λ∑kCk,ρ>0.(43)F(\mathbf{X})=\sum_k\alpha_kt_k+\lambda\sum_k C_k,\qquad\rho>0. \tag*{(43)}

The fixed-pair inner update is the following convex program:

max⁡X,{ve,k,βe,k}F(X)−ρ2∥X−X(n)∥22(44)\max_{\mathbf{X},\{v_{e,k},\beta_{e,k}\}} \quad F(\mathbf{X})-\frac{\rho}{2}\lVert\mathbf{X}-\mathbf{X}^{(n)}\rVert_2^2 \tag*{(44)}
\begin{aligned} \text{s.t.}\quad&t_k+\widetilde{R}_{e,k}(\beta_{e,k})\leq\widetilde{R}_{p,k}, &&\forall k,e,\\ &\sum_{j\in\mathcal{K}}C_j\leq\widetilde{R}_{c,k}, &&\forall k,\\ &C_k\geq r_k^{\mathrm{pub}},\quad t_k\geq r_k^{\mathrm{sec}}, &&\forall k,\\ &\lVert\mathbf{w}_c\rVert_2^2+\sum_{k\in\mathcal{K}}\lVert\mathbf{w}_k\rVert_2^2\leq P_0,\\ &\text{\xref{(34)}, \xref{(35)}}, &&\forall k,e. \tag*{(45)} \end{aligned}

Here ve,k,βe,k≥0v_{e,k},\beta_{e,k}\geq0 are imposed explicitly; Ck,tk≥0C_k,t_k\geq0 follow from nonnegative rate floors. The proximal term affects only the update; reported performance always uses FF, not the penalized objective. The problem is a convex optimization problem (maximization of a concave objective over convex constraints), and its quadratic and norm constraints admit an SOC formulation.

A positive public or secrecy floor can make an all-zero precoder initialization infeasible. For a fixed pair, use the following auxiliary Phase I problem. At iterate nn, construct the MMSE rates and leakage majorizers exactly as above and solve

min⁡X,{νe,k,βe,k},ss(46)\underset{\mathbf{X},\{\nu_{e,k},\beta_{e,k}\},s}{\min}\quad s \tag*{(46)}
s.t.tk+Le,k(n)(βe,k)≤R~p,k+s,∀k,e,(39b)∑jCj≤R~c,k+s,∀k,(39c)(38d),(38e),(38f),s≥0.(39d)\begin{aligned} \text{s.t.}\quad&t_k+\mathcal{L}^{(n)}_{e,k}(\beta_{e,k})\leq\widetilde{R}_{p,k}+s,\quad\forall k,e, &&\text{(39b)}\\ &\sum_j C_j\leq\widetilde{R}_{c,k}+s,\quad\forall k, &&\text{(39c)}\\ &(38d),\quad(38e),\\ &(38f),\quad s\geq0. &&\text{(39d)} \end{aligned}

This is convex and always admits a sufficiently large ss if the initial precoders satisfy the power bound. After every solve, make the leakage slacks tight again before forming the next majorizer. If s=0s=0, direct evaluation certifies feasibility of the conservative physical problem at that pair. If a finite initialization budget ends at s>0s>0, record “no feasible start found”; this does not prove infeasibility.

Select a finite AN design grid

Qq⊆[0,PB/N],0∈Qq.(47)\mathcal{Q}_q\subseteq[0,P_B/N],\qquad0\in\mathcal{Q}_q. \tag*{(47)}

Grid points restrict the search, not the physical AN variable. Exhaustive search evaluates at most ∣Qq∣2BM|\mathcal{Q}_q|^{2BM} pairs. For a larger RIS, a budgeted coordinate search tests alternate phases of one element and alternate AN grid values. Whenever qq or θ\theta changes, a fresh feasible start must be certified; the incumbent precoders need not remain feasible at the candidate. An outer move is accepted only if its directly evaluated conservative physical constraints are satisfied and its unpenalized objective improves by more than δout>0\delta_{\mathrm{out}}>0. Feasibility checks include the public rate floors, the exact legitimate rates, power, and (31) for every (e,k)(e,k).

Pseudocode, Complexity, and Convergence

Algorithms 1–3 specify the update and acceptance tests. In the idealized statements below, “verify” means exact evaluation; a numerical implementation must use tolerances and report its largest constraint residual.

a) Arithmetic complexity.: Let L=∣Qq∣L=|\mathcal{Q}_q| and let TIT_I and TST_S bound the Phase I and inner iterations per pair. With the cascaded channel matrices precomputed, forming effective channels for a pair takes O(MN(K+E))O(MN(K+E)) arithmetic operations; updating the legitimate MMSE quantities and evaluating all EKEK leakage amplitudes takes O(N(K2+EK))O(N(K^2+EK)) in a direct implementation. A standard SOC lift has O(NK+K2+EK)O(NK+K^2+EK) real scalar and cone coordinates; denote the actual Newton/KKT system dimension by DD and the aggregate self-concordant barrier parameter by ν\nu. For a standard interior-point solve with suitable initialization, a conservative dense factorization accounting bound is O(νD3log⁡(1/ϵip))O(\sqrt{\nu}D^3\log(1/\epsilon_{\mathrm{ip}})) arithmetic operations per solve; this excludes canonicalization and exploits no sparsity. This is a solver model, not an iteration-independent timing prediction. If HH pairs are tested, an upper accounting bound is

O(H[MN(K+E)+(TI+TS)(NK2+νD3log⁡(1/ϵip))]).(48)O\left(H\left[MN(K+E)+(T_I+T_S)\left(NK^2+\sqrt{\nu}D^3\log(1/\epsilon_{\mathrm{ip}})\right)\right]\right). \tag*{(48)}

Algorithm 1 Feasible start for a fixed (q0,θ0)(q_0,\theta_0)

Require: Pair with P0≥0P_0\geq0; total Phase I solve budget TIT_I; one or more power-feasible precoder seeds

Ensure: A verified conservative feasible X\mathbf{X}, or NotFound

  1. for each seed while total budget remains do

  2. Set Ck=rkpubC_k=r_k^{\mathrm{pub}}, tk=rksect_k=r_k^{\mathrm{sec}}; set n=0n=0

  3. while total Phase I budget remains do

  4. Set νe,k(n)\nu_{e,k}^{(n)} and βe,k(n)\beta_{e,k}^{(n)} tightly at the current precoders

  5. Update the MMSE equalizers and weights by (38)–(39)

  6. Solve Phase I (39); retain its physical variables as X(n+1)\mathbf{X}^{(n+1)}

  7. if direct evaluation certifies (31), common decoding, QoS, and power then

  8. return X(n+1)\mathbf{X}^{(n+1)}

  9. end if

  10. Increment nn and the total Phase I solve count

  11. end while

  12. end for

  13. return NotFound

Algorithm 2 Conservative fixed-pair transmit optimization

Require: Fixed feasible (q0,θ0,X(0))(q_0,\theta_0,\mathbf{X}^{(0)}); ρ>0\rho>0; maximum iterations TST_S; tolerance εS\varepsilon_S

Ensure: Verified feasible Xout\mathbf{X}^{\mathrm{out}} and its unpenalized value FoutF^{\mathrm{out}}

  1. for n=0,…,TS−1n=0,\ldots,T_S-1 do

  2. Update the MMSE auxiliaries at X(n)\mathbf{X}^{(n)} by (38)–(39)

  3. Set νe,k(n)\nu_{e,k}^{(n)} and βe,k(n)\beta_{e,k}^{(n)} tightly; form (42)

  4. Solve the SOC-representable problem (38) and obtain X(n+1)\mathbf{X}^{(n+1)}

  5. Evaluate F(X(n+1))F(\mathbf{X}^{(n+1)}) and all conservative physical constraints

  6. if verification fails then

  7. return the last verified incumbent and record a numerical failure

  8. end if

  9. if F(X(n+1))−F(X(n))≤εSF(\mathbf{X}^{(n+1)})-F(\mathbf{X}^{(n)})\leq\varepsilon_S then

  10. return (X(n+1),F(X(n+1)))(\mathbf{X}^{(n+1)},F(\mathbf{X}^{(n+1)}))

  11. end if

  12. end for

  13. return (X(TS),F(X(TS)))(\mathbf{X}^{(T_S)},F(\mathbf{X}^{(T_S)}))

Exhaustive enumeration has H=L22BMH=L^2 2^{BM}. A complete coordinate sweep tests at most M(2B−1)+(L−1)M(2^B-1)+(L-1) alternatives, although accepted moves and finite budgets determine the number of sweeps. The exponential factor in exhaustive RIS search rules out a general real-time complexity claim for that variant.

Proposition 1 (Feasibility and convergence of the inner update). Assume PB<∞P_B<\infty, all receiver-noise variances are strictly positive, and a fixed pair has a feasible starting point for the conservative physical problem. Suppose (38) is solved exactly at each iteration with ρ>0\rho>0 and the specified tight MMSE and leakage updates. Then every physical iterate is feasible for (20); F(X(n))F(\mathbf{X}^{(n)}) is nondecreasing and converges to a finite value; and ∥X(n+1)−X(n)∥2→0\lVert\mathbf{X}^{(n+1)}-\mathbf{X}^{(n)}\rVert_{2} \to0.

Proposition 2 (Verified outer acceptance). Under the same finite-power and positive-noise assumptions, suppose each tested pair is processed for a finite number of iterations and accepted only after exact physical verification and an improvement greater than δout>0\delta_{\mathrm{out}}>0. Every retained incumbent is feasible for (20), and there are finitely many accepted outer moves even if the search repeatedly revisits pairs. With a finite candidate budget, the procedure terminates. It does not assert optimality over untested or unsuccessfully initialized pairs.

Proofs are given in Appendix A. The first proposition proves objective-value convergence and vanishing physical step lengths; it does not, by itself, prove convergence of the iterates to a unique point or KKT stationarity. The outer result does not establish optimality for continuously varying qq or a globally optimal RIS phase configuration.

Simulation Results

We evaluate the robust RIS–RSMA design on one 360-kHz NR physical resource block (PRB) at 3.5 GHz. The direct BS–terminal links use the 3GPP urban-microcell (UMi) LOS/NLOS path-loss and LOS-probability expressions [18]. The RIS–terminal links use the same expressions with the RIS treated as a 10-m node, while the clear BS–RIS hop uses ITU-R P.525-5 free-space loss [19]. The latter RIS–terminal construction, independent small-scale fading across links, and independent shadowing are explicit modeling assumptions; the generator is not a spatially consistent 3GPP channel simulator. The 20-MHz carrier power is allocated uniformly to the modeled 12-subcarrier PRB at 30-kHz subcarrier spacing [20], [21]. Thus, a 30-dBm carrier budget corresponds to PB=18P_{\mathrm{B}}=18 mW (12.55 dBm) on the PRB. Table 1 gives the nominal settings; the sweep ranges are stated with the corresponding figures.

ParameterValue
Carrier frequency; carrier bandwidth; subcarrier spacing3.5 GHz (n78); 20 MHz; 30 kHz
Modeled resource; receiver noise figureOne 360-kHz PRB; 7 dB
BS antennas; served users; noncolluding eavesdroppersN=8N = 8; K=4K = 4; E=2E = 2
RIS elements and phase resolutionM=256M = 256 (16 × 16); B=2B = 2 bits
BS and RIS positions (m)(0, 0, 10); (40, 0, 10)
User positions (m)[70, 120] × [−10, 10] × {1.5}
Eavesdropper positions (m)[60, 130] × ([−35, −22] ∪ [22, 35]) × {1.5}
Conditioned direct/RIS link statesUsers: NLOS/LOS; eavesdroppers: LOS/NLOS
BS–RIS channel; RIS element efficiencyLOS free-space link; 0.8
LOS/NLOS small-scale Rician factors; shadowing standard deviations5/0; 4/7.82 dB; BS–RIS Rician factor 10
Carrier BS power; per-PRB budget PBP_{\mathrm{B}}30 dBm; 12.55 dBm (18 mW)
Per-PRB receiver noise powerσk2=σe2=−111.44\sigma_k^2 = \sigma_e^2 = -111.44 dBm
Primitive eavesdropper CSI-error fraction ρ\rho0.01; ϵd,e=ρ∥h^d,e∥2\epsilon_{d,e} = \rho\lVert\hat{\boldsymbol{h}}_{d,e} \rVert_2, ϵB,e=ρ∥B^e∥F\epsilon_{B,e} = \rho\lVert\hat{\boldsymbol{B}}_e \rVert_F
AN fraction Nq/PBN_q/P_{\mathrm{B}}{0, 0.05, 0.10, 0.20, 0.30}
Minimum public/confidential rate per userrkpub=0.1r_k^{\mathrm{pub}} = 0.1; rksec=0r_k^{\mathrm{sec}} = 0 bit/s/Hz
Objective weightsαk=1\alpha_k = 1; λ=0.2\lambda= 0.2
Screened candidates; Phase-I/inner update budgets12; 6/4
Independent seeds per operating point100 for each main case; 50 for power; 40 for other sweeps

Table 1. Nominal simulation parameters for the blocked-corridor case.

Comparison and reporting protocol: The matched controls are no AN (q=0q=0), fixed RIS (all RIS phases zero), and no RIS (the cascaded channel suppressed); the remaining variables are optimized by the same finite-budget implementation. Each plotted comparison uses common channel seeds and, except for Figure 10, conditions on the draws for which all displayed methods return directly verified solutions. We use 2000 bootstrap resamples of these paired draws for the reported 95% intervals; the sample sizes and intervals are retained in the accompanying numeric files. A failed feasible-start search is recorded as not_found, not as a certificate of physical infeasibility. The AO–SCA/SDR method of Wang et al. [5] addresses a different max–min secrecy formulation with a confidential common stream and hardware impairments. An adaptation of that method is not implemented in these experiments, so the present figures support comparisons with the three matched controls only.

Figure 10. Verified confidential goodput versus the minimum confidential rate per user; unsuccessful searches count as zero delivered service.

Confidential Rate, CSI Uncertainty, and RIS Aperture

Figure 2 shows that both methods gain from increased transmit power, while the joint design maintains a confidential-rate advantage over no RIS at every tested power. At 30 dBm, the paired-draw means are 18.95 and 16.41 bit/s/Hz, respectively; their paired difference is 2.54 bit/s/Hz with a 95% bootstrap interval of [2.03, 3.14]. The observed gain reflects the assumed reflective corridor and should not be extrapolated to arbitrary RIS placements.

Figure 2. Confidential sum rate versus carrier BS power in the blocked corridor. Power is allocated uniformly to one 360-kHz PRB.

The CSI-error fraction ρ\rho in Figure 3 affects both the direct and cascaded primitive eavesdropper channels. As ρ\rho increases from zero to 0.03, the joint confidential rate decreases from 25.27 to 12.62 bit/s/Hz. AN provides essentially no gain at ρ≤0.005\rho\leq0.005 but increases the mean by 2.86 bit/s/Hz over no AN at ρ=0.03\rho=0.03 (paired interval [1.94, 3.81]). Thus, the benefit of AN is conditional on uncertainty; it does not compensate for the overall loss due to a larger uncertainty set.

Figure 3. Confidential sum rate versus primitive eavesdropper CSI-error fraction ρ\rho. The no-AN control uses the same channel draws.

Figure 4 uses nested 8×88\times8, 12×1212\times12, and 16×1616\times16 central subarrays of the same channel draw. The joint rate increases from 16.94 to 18.89 bit/s/Hz between M=64M=64 and 256, whereas the fixed-RIS rate remains near 16.2 bit/s/Hz. The paired joint-minus-fixed-RIS gain consequently rises from

Figure 4. Confidential sum rate versus RIS aperture size MM on nested physical channel draws. Fixed RIS uses zero phases.

Algorithm 3 Budgeted AN-grid and finite-phase coordinate search

Require: Qq,QB\mathcal{Q}_{q}, \mathcal{Q}_{B}; candidate budget HH; δout>0\delta_{\mathrm{out}}>0; Phase I and inner budgets

Ensure: Best verified incumbent found, or NOTFOUND

1: Run Algorithm 1 at chosen seed pairs until a verified pair is found 2: if no verified pair is found then 3: &nbsp;&nbsp;&nbsp;&nbsp;return NOTFOUND 4: end if 5: Run Algorithm 2 at that pair; store the verified incumbent and value FincF_{\mathrm{inc}} 6: repeat 7: &nbsp;&nbsp;&nbsp;&nbsp;Set improved ←\leftarrow false 8: &nbsp;&nbsp;&nbsp;&nbsp;Form the one-element phase and alternate-qq neighbors of the current incumbent 9: &nbsp;&nbsp;&nbsp;&nbsp;Set Fbest←FincF_{\mathrm{best}}\leftarrow F_{\mathrm{inc}} and best candidate to empty 10: &nbsp;&nbsp;&nbsp;&nbsp;for each neighbor, subject to remaining budget HH do 11: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Run Algorithm 1 at the candidate pair 12: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;if the candidate has a verified feasible start then 13: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Run Algorithm 2; directly verify its physical constraints 14: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;if verification succeeds and Fcand>Fbest+δoutF_{\mathrm{cand}}>F_{\mathrm{best}}+\delta_{\mathrm{out}} then 15: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Store this candidate as best; set Fbest←FcandF_{\mathrm{best}}\leftarrow F_{\mathrm{cand}} 16: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;end if 17: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;end if 18: &nbsp;&nbsp;&nbsp;&nbsp;end for 19: &nbsp;&nbsp;&nbsp;&nbsp;if a best candidate was stored then 20: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Replace the incumbent; set Finc←FbestF_{\mathrm{inc}}\leftarrow F_{\mathrm{best}} and improved ←\leftarrow true 21: &nbsp;&nbsp;&nbsp;&nbsp;end if 22: until improved = false or budget HH is exhausted 23: return the verified incumbent with public and confidential rates reported separately 0.70 to 2.66 bit/s/Hz. In this geometry, a larger aperture is useful primarily when its phases are adapted, rather than merely increased in number.

Public Traffic, Rate Distribution, and AN Selection

Figure 5 varies the public-rate weight over λ∈{0,0.1,0.2,0.5,1}\lambda\in\{0, 0.1, 0.2, 0.5, 1\}. For the joint design, increasing the delivered public sum rate from 0.40 to 1.07 bit/s/Hz changes the confidential sum rate from 18.89 to 18.57 bit/s/Hz. The common stream contributes only to the horizontal coordinate: its public bits are excluded from confidential throughput. The limited movement of the points at intermediate weights is an outcome of the finite candidate search, rather than a claim that the plotted points trace a globally optimal rate region.

Figure 5. Delivered public common rate and confidential private rate as the public-rate weight λ\lambda varies.

The empirical distributions in Figure 6 use the intersection of 97 verified nominal draws across all four methods. The paired mean gain of the joint design is 2.59 bit/s/Hz over no RIS and 2.78 bit/s/Hz over fixed RIS, with respective 95% intervals [2.10, 3.12] and [2.22, 3.41]. Its mean gain over no AN is only 0.25 bit/s/Hz and occurs in a small subset of draws; the joint design exceeds no AN on 7 of the 97 paired draws. This distributional view avoids interpreting a small average AN gain as a typical per-channel gain.

Figure 6. Empirical CDF of confidential sum rate on the 97 nominal draws verified by all four methods.

Figure 7 explains this behavior: of the 99 verified nominal joint-design draws, 92 select zero AN and 7 select Nq/PB=0.05N_q/P_{\mathrm{B}} = 0.05; none select a larger fraction. At ρ=0.03\rho= 0.03 in the uncertainty sweep, positive AN is instead selected on 23 of 39 verified draws. Hence, the AN grid is available to the optimizer, but its use is selective under the stated conservative leakage bound.

Figure 7. Selection frequency of the total AN power fraction Nq/PBN_q/P_B on 99 verified nominal draws.

Algorithm Behavior and Propagation Sensitivity

Figure 8 reports one fixed-pair diagnostic: zero RIS phases, Nq/PB=0.05N_q/P_B = 0.05, and the prespecified median-rate nominal channel draw. The directly evaluated, unpenalized objective rises from 11.14 to 14.54 bit/s/Hz and levels off within approximately ten inner updates. This diagnostic uses a 20-update budget, whereas the Monte Carlo campaign uses four inner updates; it neither depicts full alternating RIS/AN convergence nor establishes finite-run global optimality.

Figure 8. Direct objective trace for a single fixed zero-phase RIS and Nq/PB=0.05N_q/P_B = 0.05 pair.

Figure 9 gives a paired joint-minus-no-RIS gain of 1.40 bit/s/Hz under random UMi LOS states and 2.59 bit/s/Hz in the conditioned blocked corridor. The respective bootstrap intervals are [1.03, 1.80] and [2.10, 3.12]. The two cases also differ in terminal geometry and link-state conditioning; this

Figure 9. Paired confidential-rate gain of joint RIS/AN over no RIS in the random-LOS and conditioned blocked-corridor cases; whiskers show 95% bootstrap intervals.

comparison therefore measures the gain within each specified case, not the isolated causal effect of blockage.

Fig. 10} imposes per-user minimum confidential rates from 0 to 1.5 bit/s/Hz. Here the ordinate is verified confidential goodput: a search that does not return a verified design contributes zero service. At the largest requirement, the joint and no-RIS goodputs are 13.77 and 12.23 bit/s/Hz, while verified-solution counts fall to 30/40 and 29/40, respectively. The decline therefore reflects both the tighter service requirement and the reduced rate of successful feasible-start searches; it must not be described as a measured physical infeasibility probability.

Search Time and Phase Resolution

Figure 11 reports wall-clock time for the finite-budget implementation. The joint-design mean rises from 1.27 to 1.45 as MM increases from 64 to 256, but the fixed-RIS curve is nonmonotonic (1.69, 1.80, and 1.38 s). Variable solver paths and runtime variability preclude fitting an asymptotic complexity law to these three operating points; the values characterize this implementation and computing environment only.

Figure 11. Measured finite-budget search time versus RIS size for the joint, fixed-RIS, and no-RIS designs.

Finally, Figure 12 shows that increasing the phase resolution from one to two bits raises the joint confidential rate from 18.35 to 18.89 bit/s/Hz; four bits yield 19.06 bit/s/Hz. The fixed-zero-phase control is independent of BB. Thus, two-bit phases capture most of the observed four-bit performance under this channel generator and candidate budget, without establishing a universal hardware-resolution optimum.

Figure 12. Confidential sum rate versus RIS phase resolution BB on paired physical channel draws.

Conclusion

This paper formulated a robust RIS-assisted RSMA downlink for mixed public and confidential traffic. The common stream carries public information, while confidential information is conveyed through private streams. The secrecy model allows each eavesdropper to cancel the common stream and treats the other private streams as known side information. Bounded errors in the direct and cascaded eavesdropper channels lead to an effective-channel uncertainty bound that remains valid as the RIS phases change.

We developed a conservative joint design of the transmit precoders, artificial noise, common-rate allocations, and finite-resolution RIS phases. The procedure combines feasible initialization, convex inner transmit updates, and a discrete outer search with direct feasibility checks. For a fixed artificial-noise power and RIS configuration, feasible inner iterates have a nondecreasing objective that converges in value; accepted outer updates remain feasible and are finite in number. These guarantees establish what the procedure delivers under the stated uncertainty model without implying global optimality.

Appendix A: Proofs of the Algorithmic Guarantees

Proof of Proposition 1. At iteration nn, choose the MMSE auxiliaries at the physical incumbent. The resulting rate lower bounds equal the exact legitimate rates at X(n)\mathbf{X}^{(n)}. Choose the leakage slacks at their smallest feasible values: then βe,k(n)=(∣h^eHwk∣+ϵe∥wk∥2)2/de\beta_{e,k}^{(n)} = (|\hat{\mathbf{h}}_{e}^{H}\mathbf{w}_{k}|+\epsilon_{e}\lVert\mathbf{w}_{k}\rVert_{2})^{2}/d_{e}. The affine leakage bound equals the conservative leakage rate at these slacks. Consequently, (X(n),v(n),β(n))(\mathbf{X}^{(n)},\mathbf{v}^{(n)},\boldsymbol{\beta}^{(n)}) is feasible for (38) whenever the physical incumbent is feasible for (31), common decoding, QoS, and power.

For any feasible point of (38), R‾e,kub≤log⁡2(1+βe,k)≤Le,k(n)(βe,k)\overline{R}_{e,k}^{\mathrm{ub}}\le\log_{2}(1+\beta_{e,k})\le\mathcal{L}_{e,k}^{(n)}(\beta_{e,k}), whereas R~p,k≤Rp,k\widetilde{R}_{p,k}\le R_{p,k} and R~c,k≤Rc,k\widetilde{R}_{c,k}\le R_{c,k}. Hence, its physical variables satisfy (31) and the exact common, QoS, and power constraints. By induction, all iterates are feasible for the conservative physical problem and therefore also for (20).

Use the old incumbent and tight slacks as a feasible point in the optimization at iteration nn. Optimality of the new point gives

F(X(n+1))−F(X(n))≥ρ2∥X(n+1)−X(n)∥22≥0.(49)F(\mathbf{X}^{(n+1)})-F(\mathbf{X}^{(n)})\ge\frac{\rho}{2}\lVert\mathbf{X}^{(n+1)}-\mathbf{X}^{(n)}\rVert_{2}^{2}\ge0. \tag*{(49)}

Because RIS phases belong to a finite alphabet, each Hk≜max⁡θ∈QMB∥hk(θ)∥22H_k\triangleq\max_{\boldsymbol{\theta}\in\mathcal{Q}_{M}^{B}}\lVert\mathbf{h}_{k}(\boldsymbol{\theta})\rVert_{2}^{2} is finite. The power budget and positive noise give the uniform bound Rc,k,Rp,k≤log⁡2(1+PBHk2/σk2)R_{c,k},R_{p,k}\le\log_{2}(1+P_{B}H_{k}^{2}/\sigma_{k}^{2}). The secrecy inequalities imply 0≤tk≤Rp,k0\le t_k\le R_{p,k}; common decoding gives ∑kCk≤max⁡iRc,i\sum_k C_k\le\max_i R_{c,i}. Thus FF has a finite upper bound. A nondecreasing bounded sequence of real values converges. Summing (49) over nn bounds ∑n≥0∥X(n+1)−X(n)∥22\sum_{n\ge0}\lVert\mathbf{X}^{(n+1)}-\mathbf{X}^{(n)}\rVert_{2}^{2} by a finite constant, proving that its individual terms tend to zero.

Proof of Proposition 2. Algorithm 3 stores only physically verified designs. Proposition 1 and the direct candidate check imply that every accepted incumbent satisfies the conservative constraints and hence the original robust constraints. Let Fmax⁡<∞F_{\max}<\infty be the uniform objective bound from the preceding proof and F0F_0 be the initial incumbent value. After JJ accepted moves, FJ>F0+JδoutF_J>F_0+J\delta_{\mathrm{out}}; therefore J<(Fmax⁡−F0)/δoutJ<(F_{\max}-F_0)/\delta_{\mathrm{out}}. Each sweep checks a finite set of alternatives and each candidate uses finite Phase I and inner budgets. Thus the number of accepted moves is finite, and a finite overall candidate budget also forces termination. A failed Phase I run conveys no infeasibility certificate and cannot be used to infer optimality.

References

References

  1. [1]X. Yu, D. Xu, Y. Sun, D. W. K. Ng, and R. Schober, “Robust and secure wireless communications via intelligent reflecting surfaces,” IEEE Journal on Selected Areas in Communications, vol. 38, no. 11, pp. 2637–2652, 2020.arxiv.org/abs/1912.01497
  2. [2]Z. Peng, R. Weng, C. Pan, G. Zhou, M. D. Renzo, and A. L. Swindlehurst, “Robust transmission design for RIS-assisted secure multiuser communication systems in the presence of hardware impairments,” IEEE Transactions on Wireless Communications, vol. 22, no. 11, pp. 7506–7521, 2023.arxiv.org/abs/2202.11860
  3. [3]Q. Wu and R. Zhang, “Beamforming optimization for wireless network aided by intelligent reflecting surface with discrete phase shifts,” IEEE Transactions on Communications, vol. 68, no. 3, pp. 1838–1851, 2020.arxiv.org/abs/1906.03165
  4. [4]Y. Mao, B. Clerckx, and V. O. K. Li, “Rate-splitting for multi-antenna non-orthogonal unicast and multicast transmission: Spectral and energy efficiency analysis,” IEEE Transactions on Communications, vol. 67, no. 12, pp. 8754–8770, 2019.arxiv.org/abs/1808.08325
  5. [5]K. Wang, D. Li, J. Zhang, and C. Li, “Robust and secure transmission for RIS-assisted RSMA with transceiver hardware impairments,” Physical Communication, vol. 71, p. 102629, 2025. [Online]. Available: https://doi.org/10.1016/j.phycom.2025.102629
  6. [6]A. B. M. Adam, M. A. Ouamri, X. Wan, M. S. A. Muthanna, R. Alkanhel, A. Muthanna, and X. Li, “Secure communication in uav–ris-empowered multiuser networks: Joint beamforming, phase shift, and uav trajectory optimization,” IEEE Systems Journal, vol. 18, no. 2, pp. 1009–1019, 2024.DOI
  7. [7]A. B. M. Adam, X. Wan, M. A. M. Elhassan, M. S. A. Muthanna, A. Muthanna, N. Kumar, and M. Guizani, “Intelligent and robust uav-aided multiuser ris communication technique with jittering uav and imperfect hardware constraints,” IEEE Transactions on Vehicular Technology, vol. 72, no. 8, pp. 10 737–10 753, 2023.DOI
  8. [8]A. Kakati, G. Li, E. M. Diallo, M. A. Ouamri, M. I. Khalid, A. B. M. Adam, and X. Li, “Hybrid ris-assisted multi-user indoor-outdoor secure communication in aav mmwave networks,” IEEE Transactions on Vehicular Technology, vol. 75, no. 6, pp. 10 114–10 129, 2026.DOI
  9. [9]E. M. Diallo, A. B. M. Adam, A. Kakati, M. A. Diallo, and M. Amine Ouamri, “Deep learning-based secure content fetching in uav-aided and ris-assisted under coordinated aerial and ground eavesdropping,” in 2025 IEEE International Conference on Communications Workshops (ICC Workshops), 2025, pp. 1544–1549.DOI
  10. [10]A. B. M. Adam and M. A. M. Elhassan, “Enhancing secrecy in uav rsma networks: Deep unfolding meets deep reinforcement learning,” in 2023 International Conference on Communications, Computing, Cyber-security, and Informatics (CCCI), 2023, pp. 1–5.
  11. [11]M. A. Ouamri, A. B. M. Adam, Y. Benallouche, and M. Gueroui, “Secure qoe-aware uav-aided rate-splitting multiple access-based communications,” in GLOBECOM 2025 - 2025 IEEE Global Communications Conference, 2025, pp. 4565–4570.DOI
  12. [12]M. A. Ouamri, A. B. M. Adam, and M. A. M. Elhassan, “Quantum deep reinforcement learning for secure multi-user uav rsma networks,” in 2025 12th IFIP International Conference on New Technologies, Mobility and Security (NTMS), 2025, pp. 86–92.DOI
  13. [13]A. B. M. Adam, M. A. Ouamri, M. S. A. Muthanna, X. Li, M. A. M. Elhassan, and A. Muthanna, “Real-time and security-aware precoding in ris-empowered multi-user wireless networks,” 2023. [Online]. Available: https://arxiv.org/abs/2401.00192
  14. [14]A. B. M. Adam, Z. Abdullah, and S. Chatzinotas, “Generative ai-based hierarchical drl framework for ris-assisted thz massive mimo systems,” IEEE Transactions on Wireless Communications, vol. 25, pp. 4480–4495, 2026.DOI
  15. [15]M. A. Diallo, M. A. Ouamri, M. I. Khalid, A. Komal, E. M. Diallo, and A. B. M. Adam, “Multi-modal deep reinforcement learning framework for interference-limited ris-assisted uav wireless network,” in 2025 12th IFIP International Conference on New Technologies, Mobility and Security (NTMS), 2025, pp. 80–85.DOI
  16. [16]M. A. Ouamri, A. B. M. Adam, M. A. M. Elhassan, E. Moustapha Diallo, and X. Li, “Discrete softmax policy gradient for statistical qos provisioning in ris-aided holographic mimo networks,” IEEE Communications Letters, vol. 29, no. 11, pp. 2516–2520, 2025.DOI
  17. [17]E. M. Diallo, A. B. M. Adam, M. A. M. Elhassan, M. Hammoudeh, M. S. A. Muthanna, and A. Muthanna, “Quantum-enhanced multi-ris-empowered consumer electronics mmwave iot noma networks,” IEEE Transactions on Consumer Electronics, vol. 72, no. 3, pp. 5965–5973, 2026.DOI
  18. [18]3GPP, “TR 38.901: Study on channel model for frequencies from 0.5 to 100 GHz,” 3rd Generation Partnership Project, Tech. Rep. V19.5.0, Sep. 2026. [Online]. Available: https://www.etsi.org/deliver/etsi_tr/138900_138999/138901/19.05.00_60/tr_138901v190500p.pdf
  19. [19]ITU-R, “Calculation of free-space attenuation,” International Telecommunication Union, Tech. Rep. Recommendation ITU-R P.525-5, Nov. 2024. [Online]. Available: https://www.itu.int/rec/R-REC-P.525
  20. [20]3GPP, “TS 38.211: NR; physical channels and modulation,” 3rd Generation Partnership Project, Tech. Rep. V18.8.0, Feb. 2026. [Online]. Available: https://www.etsi.org/deliver/etsi_ts/138200_138299/138211/18.08.00_60/ts_138211v180800p.pdf
  21. [21]3GPP, “TS 38.104: NR; base station (BS) radio transmission and reception,” 3rd Generation Partnership Project, Tech. Rep. V19.3.0, Feb. 2026. [Online]. Available: https://www.etsi.org/deliver/etsi_ts/138100_138199/138104/19.03.00_60/ts_138104v190300p.pdf

Paper details

vlnc:W_65wkkknk
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents